CN113520683B - Lower limb artificial limb control system and method based on imitation learning - Google Patents

Abstract

The invention provides a lower limb prosthesis control system and method based on imitation learning, and relates to the technical field of control strategies for lower limb prostheses. The system includes a knee joint control motor, an ankle joint control motor, a connecting rod and a casing. The method includes: a data collection step: a healthy human body wears an IMU sensor, collects gait information of a healthy human body in different stances, and forms a training data set; a simulation training model step: builds a convolutional neural network model, and uses the collected information in a simulation environment The data of the model is trained; the entity verification step: transplant the convolutional neural network model to the lower limb prosthetic controller, when the lower limb prosthetic controller receives the signal input from the IMU sensor, use the trained convolutional neural network model to output the lower limb The action command of the prosthetic joint realizes the control of the lower limb prosthesis. The invention can realize end-to-end control, and the control process is smoother, and at the same time, the cumbersome procedure of labeling the training data set is reduced.

Description

Control system and method for lower limb prosthesis based on imitation learning

technical field

The invention relates to the technical field of control strategies for lower limb prostheses, in particular to a control system and method for lower limb prostheses based on imitation learning.

Background technique

With the improvement of social living standards, lower limb prostheses have been more and more widely used in recent years. But there is room for further improvement in how lower-limb prosthetics are controlled. Imitation learning algorithms are now widely used in many fields of society, such as teaching robots, automatic driving, etc.

The invention patent with the publication number CN103750927B discloses a method for controlling the motion of the knee joint of a human lower limb prosthesis, in particular an adaptive iterative learning control method for a two-rigid-body two-degree-of-freedom lower-limb prosthesis walking on flat ground. The invention firstly analyzes the normal gait characteristics of the human body and the control requirements of the lower limb prosthesis; then performs dynamic analysis on it through the Newton-Euler algorithm, and establishes a two-rigid body two-degree-of-freedom lower limb prosthesis motion system model; finally applies the self-adaptive iterative learning control algorithm For this motion system model, its control algorithm process includes: problem description, convergence analysis, boundedness of seeking, incrementality of seeking, and increasing of seeking.

In view of the above prior art, the existing lower limb prosthetic control scheme cannot achieve good end-to-end control, and the procedure in the process of labeling the training data set is cumbersome.

Contents of the invention

Aiming at the defects in the prior art, the present invention provides a lower limb prosthesis control system and method based on imitation learning.

According to a lower limb prosthesis control system and method based on imitation learning provided by the present invention, the scheme is as follows:

In the first aspect, a lower limb prosthesis control system based on imitation learning is provided, the system includes a lower limb prosthesis entity, and the lower limb prosthesis entity includes: a knee joint control motor, an ankle joint control motor, a connecting rod and a housing; The two ends of the rod are respectively connected to the knee joint control motor and the ankle joint control motor, and the shell is arranged on the circumferential image side wall of the connecting rod.

Preferably, multiple IMU sensors are mounted on the lower limb prosthesis body.

In a second aspect, a method for controlling a lower limb prosthesis based on imitation learning is provided, the method comprising:

Data collection steps: healthy people wear IMU sensors, perform standard gait demonstrations in different scenarios, collect gait information of healthy people in different gaits, and form training data sets;

Simulation training model steps: build a convolutional neural network model, and use the collected data to train the model in the simulation environment;

Entity verification step: transplant the convolutional neural network model to the lower limb prosthetic physical controller, when the lower limb prosthetic controller receives the signal input from the IMU sensor, use the trained convolutional neural network model to output the movement instructions of the lower limb prosthetic joints to realize Control of lower limb prostheses.

Preferably, the training data in the step of collecting data is the angle and angular velocity information of hip joints, knee joints, and ankle joints of a healthy human body in different stances.

Preferably, the training data collection process is:

An IMU sensor is respectively fixed at the instep, calf and thigh of the lower limb prosthetic body;

The angle and angular velocity of the ankle joint are obtained by subtracting the IMU sensor readings at the instep position from the IMU sensor readings at the calf position;

The angle and angular velocity of the knee joint are obtained by subtracting the IMU sensor reading of the thigh position from the IMU sensor reading of the calf position;

The angle and angular velocity of the hip joint are obtained from the thigh position IMU sensor readings.

Preferably, the angle and angular velocity information is input into the convolutional neural network model through a dimensionally stretched matrix of 256*256, and the convolutional neural network model has 3 convolutional layers, each convolutional layer is followed by a maximum pooling layer , two fully connected layers are connected after the convolutional layer and the maximum pooling layer, and the output layer is connected after the fully connected layer. The activation function used in the convolutional neural network is the linear rectification function ReLU;

The output of the convolutional neural network model is the current value of the knee joint control motor and the ankle joint control motor.

Preferably, the training of the convolutional network model is completed in a simulation environment, and the angle and angular velocity information is subjected to data standardization and data enhancement processing before being input into the convolutional neural network.

Preferably, the data standardization method is to subtract the angle and angular velocity data obtained when the human body stands still from the real-time measurement data; the data enhancement is completed by adding noise data and simulation data.

Preferably, a convolutional neural network model is established to observe healthy human gait data, and the mathematical description is expressed as:

Among them, d(θ) represents the characteristic distribution of healthy human gait, p(θ) represents the characteristic distribution of predicted actions, π′ is the optimal strategy for imitation learning, S(d(θ),p(θ)) is Comparison of similarities between the two.

Preferably, the loss function in the training process of the convolutional neural network model is set as follows:

Among them, A _h1 represents the angle of the ankle joint of the lower limb of the healthy person, A _p1 represents the angle of the ankle joint of the lower limb prosthesis, K _h1 represents the angle of the knee joint of the lower limb of the healthy person, K _p1 represents the angle of the knee joint of the lower limb prosthesis, H _h1 represents the angle of the hip joint of the lower limb of the healthy person, H _p1 represents the angle of the hip joint of the lower limb prosthesis, A _h2 represents the angular velocity of the ankle joint of the lower limb of the healthy person, A _p2 represents the angular velocity of the ankle joint of the lower limb prosthesis, K _h2 represents the angular velocity of the knee joint of the lower limb of the healthy person, K _p2 represents the angular velocity of the knee joint of the lower limb prosthesis, and H _h2 represents The angular velocity of the hip joint of the lower limb of the healthy human body, H _p2 represents the angular velocity of the hip joint of the lower limb prosthesis.

Compared with the prior art, the present invention has the following beneficial effects:

1. The control strategy of the present invention realizes end-to-end control, eliminates the cumbersome modeling process in the middle, makes the control process of the lower limb prosthesis more smooth and concise, and reduces the cumbersome procedures of labeling the training data set;

2. The gait data of healthy human beings collected in the present invention are not directly put into the network model for training, but the robustness and generalization of the model are increased through standardization and data enhancement.

Description of drawings

Other characteristics, objects and advantages of the present invention will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings:

Fig. 1 is the main structural diagram of the lower limb prosthetic model;

Figure 2 is a schematic diagram of the fixed position of the IMU sensor;

3 is a schematic diagram of the overall flow of the lower limb prosthetic control method;

Fig. 4 is a schematic diagram of the algorithm framework of behavior cloning imitation learning;

Fig. 5 is a schematic diagram of the training process of the convolutional neural network model.

Reference signs: 1. Knee joint control motor; 2. Ankle joint control motor; 3. Connecting rod; 4. Shell; 5. IMU sensor.

Detailed ways

The present invention will be described in detail below in conjunction with specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

The embodiment of the present invention provides a lower limb prosthesis control system based on imitation learning, as shown in FIG. 1 and FIG. The connecting rod 3 and the shell 4; the two ends of the connecting rod 3 are respectively connected to the knee joint control motor 1 and the ankle joint control motor 2, and the shell 4 is wrapped and arranged on the circumferential image side wall of the connecting rod 3, and a lower limb prosthesis body is equipped with three One IMU sensor 5, and two legs fix six IMU sensors 5 in total.

This embodiment also provides a lower limb prosthetic control method based on imitation learning. Referring to FIG. 3 and FIG. 4, the present invention collects gait data of a healthy human body and establishes a convolutional neural network model to observe the gait data of a healthy human body. Find a strategy to reproduce the gait of a healthy human body as much as possible, and its mathematical description can be expressed as:

Among them, d(θ) represents the characteristic distribution of healthy human gait, p(θ) represents the characteristic distribution of predicted actions, π′ is the optimal strategy for imitation learning, S(d(θ),p(θ)) is Comparison of similarities between the two.

More specifically, the imitation learning method adopted in this embodiment is mainly a behavior cloning method, which essentially uses the gait data of a healthy human body for supervised learning. In this embodiment, the gait data of a healthy human body is divided into left leg data and right leg data. During the training process, the left leg data is the state data Si, and the right leg data is the fitting target data.

In this embodiment, the convolutional neural network model does not directly output the gait information of the right leg to achieve imitation learning, but outputs the current values of the lower limb prosthesis knee joint control motor 1 and ankle joint control motor 2, and controls the motor through the knee joint. 1 and the ankle joint control motor 2 drive the knee joint and ankle joint of the lower limb prosthesis to move, and obtain the readings of the IMU sensor 5 (Inertial Measurement Unit inertial sensor) on the lower limb prosthesis after the action, according to the readings of the IMU sensor 5 on the lower limb prosthesis and the healthy human body Strategy selection based on how well the right leg sensor readings fit.

Specifically, the gait data of a healthy human body includes the angle and angular velocity information of the hip joint, knee joint, and ankle joint of the left leg and the angle and angular velocity information of the hip joint, knee joint, and ankle joint of the right leg under different stances of the healthy human body. The acquisition process is as follows: one IMU sensor 5 is respectively fixed on the instep, the calf and the thigh, and a total of six IMU sensors 5 are fixed on the two legs. Among them, the angle and angular velocity of the ankle joint are obtained by subtracting the readings of the calf position IMU sensor 5 from the readings of the instep position IMU sensor 5, the angle and angular velocity of the knee joint are obtained by subtracting the readings of the thigh position IMU sensor 5 from the readings of the calf position IMU sensor 5, and the hip joint The angle and angular velocity of are obtained by the thigh position IMU sensor 5 readings.

The obtained healthy human gait data is not directly put into the network model for training, but to increase the robustness and generalization of the model through standardization and data enhancement. The angle and angular velocity data obtained when not moving, the data enhancement is completed by adding a small amount of noise data and a small amount of simulation data.

The preprocessed data is input into the convolutional neural network through a matrix of dimensions stretched to 256*256. The convolutional neural network has 3 convolutional layers, and each convolutional layer is followed by a maximum pooling layer. In the convolutional layer And the maximum pooling layer is followed by two fully connected layers, and the fully connected layer is followed by the output layer. The activation function used in the convolutional neural network is the linear rectification function ReLU (Rectified Linear Unit).

Specifically, the loss function in the training process of the convolutional neural network model is set as follows:

Among them, A _h1 represents the angle of the ankle joint of the lower limb of the healthy person, A _p1 represents the angle of the ankle joint of the lower limb prosthesis, K _h1 represents the angle of the knee joint of the lower limb of the healthy person, K _p1 represents the angle of the knee joint of the lower limb prosthesis, H _h1 represents the angle of the hip joint of the lower limb of the healthy person, H _p1 represents the angle of the hip joint of the lower limb prosthesis, A _h2 represents the angular velocity of the ankle joint of the lower limb of the healthy person, A _p2 represents the angular velocity of the ankle joint of the lower limb prosthesis, K _h2 represents the angular velocity of the knee joint of the lower limb of the healthy person, K _p2 represents the angular velocity of the knee joint of the lower limb prosthesis, and H _h2 represents The angular velocity of the hip joint of the lower limb of the healthy human body, H _p2 represents the angular velocity of the hip joint of the lower limb prosthesis.

Referring to Figure 5, considering safety and training efficiency, the convolutional neural network model is first trained in the simulation environment, and after the training converges, the entity is transplanted, and then the convolutional neural network model is further targeted according to the control effect of the entity. sex training. The entity and the simulation iterate each other, and the iteration stops after the network model achieves the expected control effect. The example is as follows, the convolutional neural network model trained and converged in the simulation environment is transplanted to the entity, and it is found that the convolutional neural network model controls the lower limb prosthetic entity to slip when going up and down the stairs. At this time, the gait of healthy people going up and down the stairs The data is specially collected, and then the convolutional neural network model is trained in the simulation environment with the collected gait data of healthy people going up and down the stairs. After the training results reach the expected effect, the model is transplanted to the lower limb prosthetic entity again. Verification, if the control defect of the convolutional neural network model is found in the verification of the lower limb prosthesis again, the corresponding data will be specially collected again to conduct targeted training on the convolutional neural network model, and so on until the convolutional neural network model is satisfactory. control effect.

After wearing the lower limb prosthesis controlled by the strategy of the present invention, the disabled only need to fix three IMU sensors 5 at the instep position, thigh position, and calf position of the prosthesis and the instep position, thigh position, and calf position of the healthy leg. The specific fixed positions are as follows: As shown in Figure 2, when the healthy leg moves, the lower limb prosthesis will also move accordingly. Compared with the traditional lower limb prosthesis control strategy, the control strategy of the present invention realizes end-to-end control, eliminates the cumbersome modeling process in the middle, and makes the control process of the lower limb prosthesis more smooth and concise, and subtracts the labeling training data set cumbersome procedures.

It should be noted that the convolutional neural network model of the present invention may not be limited to the specific input in the specification. For example, in order to increase the intention recognition rate of the convolutional neural network model for lower limb prosthesis wearers, the surrounding environment image information can be added to the training The convolutional neural network model is trained in the data set. After the training is completed, the surrounding environment image information should also be added when the lower limb prosthetic entity is controlled. According to needs, EMG signals, EEG signals and other information can also be used as the input of the convolutional neural network model. Those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention.

The embodiment of the present invention provides a lower limb prosthesis control system and method based on imitation learning. The control strategy of the present invention realizes end-to-end control, eliminating the cumbersome modeling process in the middle, and making the control process of the lower limb prosthesis more smooth and concise. At the same time, the tedious procedure of labeling the training data set is reduced.

Those skilled in the art know that, in addition to realizing the system provided by the present invention and its various devices, modules, and units in a purely computer-readable program code mode, the system provided by the present invention and its various devices can be completely programmed by logically programming the method steps. , modules, and units implement the same functions in the form of logic gates, switches, ASICs, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by the present invention can be regarded as a hardware component, and the devices, modules, and units included in it for realizing various functions can also be regarded as hardware components. The structure; the devices, modules, and units for realizing various functions can also be regarded as not only the software modules for realizing the method, but also the structures in the hardware components.

Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. In the case of no conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims (8)

1. A lower limb prosthesis control method based on imitation learning, it is characterized in that, the lower limb prosthesis control system based on imitation learning that adopts comprises: lower limb prosthesis entity, and described lower limb prosthesis entity comprises: knee joint control motor (1), ankle joint Control motor (2), connecting rod (3) and shell (4); the two ends of described connecting rod (3) are respectively connected to knee joint control motor (1) and ankle joint control motor (2), and described shell ( 4) be arranged on the circumferential image side wall of the connecting rod (3); A plurality of IMU sensors (5) are installed on the lower limb prosthesis entity; include: Data collection step: healthy people wear IMU sensors (5), perform standard gait demonstrations in different scenarios, collect gait information of healthy people in different gaits, and form training data sets; Simulation training model steps: build a convolutional neural network model, and use the collected data to train the model in the simulation environment; Entity verification step: transplant the convolutional neural network model to the lower limb prosthesis physical controller, when the lower limb prosthesis controller receives the signal input from the IMU sensor (5), use the trained convolutional neural network model to output the movement of the lower limb prosthetic joint command to realize the control of the lower limb prosthesis. 2. the lower limb prosthesis control method based on imitation learning according to claim 1, is characterized in that, the training data in the described data collection step is the angle of hip joint, knee joint, ankle joint of healthy human body under different stances and angular velocity information. 3. the lower limb prosthesis control method based on imitation learning according to claim 2, is characterized in that, described training data acquisition process is: An IMU sensor (5) is respectively fixed at the instep, calf and thigh of the lower limb prosthesis entity; The angle and angular velocity of the ankle joint are obtained by subtracting the readings of the calf position IMU sensor (5) from the readings of the instep position IMU sensor (5); The angle and angular velocity of the knee joint are obtained by subtracting the reading of the thigh position IMU sensor (5) from the reading of the calf position IMU sensor (5); The angle and angular velocity of the hip joint are read by the thigh position IMU sensor (5). 4. The method for controlling lower limb prostheses based on imitation learning according to claim 2, wherein the angle and angular velocity information is stretched into a 256*256 matrix input convolutional neural network model, and the convolutional neural network model There are 3 convolutional layers, each convolutional layer is followed by a maximum pooling layer, two fully connected layers are connected after the convolutional layer and the maximum pooling layer, and the fully connected layer is followed by the output layer. In the convolutional neural network The activation function used is the linear rectification function ReLU; The output of the convolutional neural network model is the current value of the knee joint control motor (1) and the ankle joint control motor (2). 5. The lower limb prosthesis control method based on imitation learning according to claim 4, wherein the training of the convolutional network model is completed in a simulation environment, and angle and angular velocity information are carried out before inputting the convolutional neural network Data standardization and data enhancement processing. 6. the lower limb prosthesis control method based on imitation learning according to claim 5, is characterized in that, described data normalization mode is to subtract the angle and the angular velocity data that human body obtains when standing still with real-time measurement data; Data enhancement then This is done by adding noise data as well as simulated data. 7. the lower limb prosthesis control method based on imitation learning according to claim 1, is characterized in that, set up convolutional neural network model and observe healthy human body gait data, mathematical description is expressed as:

Among them, d(θ) represents the characteristic distribution of healthy human gait, p(θ) represents the characteristic distribution of predicted actions, π′ is the optimal strategy for imitation learning, S(d(θ),p(θ)) is Comparison of similarities between the two. 8. the lower limb prosthesis control method based on imitation learning according to claim 1, is characterized in that, the loss function in described convolutional neural network model training process is set as follows:

Among them, A _h1 represents the angle of the ankle joint of the lower limb of the healthy person, A _p1 represents the angle of the ankle joint of the lower limb prosthesis, K _h1 represents the angle of the knee joint of the lower limb of the healthy person, K _p1 represents the angle of the knee joint of the lower limb prosthesis, H _h1 represents the angle of the hip joint of the lower limb of the healthy person, H _p1 represents the angle of the hip joint of the lower limb prosthesis, A _h2 represents the angular velocity of the ankle joint of the lower limb of the healthy person, A _p2 represents the angular velocity of the ankle joint of the lower limb prosthesis, K _h2 represents the angular velocity of the knee joint of the lower limb of the healthy person, K _p2 represents the angular velocity of the knee joint of the lower limb prosthesis, and H _h2 represents The angular velocity of the hip joint of the lower limb of the healthy human body, H _p2 represents the angular velocity of the hip joint of the lower limb prosthesis.

Images